Skip to main content
Observability is one of the layers TrueFoundry adds on top of the open-source TrueForge harness — see What TrueFoundry Adds for the rest. Every agent session runs through the AI Gateway, so each turn, tool call, and model call is captured with cost, tokens, and latency, with no extra instrumentation in your agent. That gives you one place to answer three questions: how an agent is being used, how much each run costs, and where a run went wrong.

​
Agent sessions

Sessions in the TrueForge portal lists every session across all your agents. Each entry is one conversation, labelled with its opening message, the agent it ran against, and its turn count, cost, and duration — so you can scan for expensive or unusually long runs without opening anything. Sessions can be narrowed to a single agent or to a time window.
Agent Sessions view in the TrueForge portal listing sessions across agents, each showing the opening message, the agent it ran against, how long ago it ran, and its turn count, cost, and duration, with filters for agent and time range

Sessions lists every conversation across all agents with its agent, turn count, cost, and duration

​
Session detail

Opening a session shows the whole run: aggregate metrics at the top, a timeline of what the harness was doing, and the full transcript below.
Agent session detail showing summary metrics for turns, duration, cost, tokens, context, tool calls, sub-agents and errors, an event-type timeline coloured by system, user, model and tool call, and the turn-by-turn transcript with an expanded Agent steps panel listing an MCP tool call

Session detail shows aggregate metrics, the event-type timeline, and the turn-by-turn transcript with expandable agent steps

The summary metrics cover the session as a whole: Beneath them, the event-type timeline breaks the run into system, user, model, and tool-call segments across each turn, which makes it obvious where the time actually went — a turn dominated by orange is tool-bound, one dominated by blue is model-bound. The transcript carries its own tokens, duration, and cost per turn, and every turn’s agent steps can be expanded to see the individual tool calls the harness made, in order, with their outcome. That is usually enough to find where a run went wrong: a cancelled turn, a failing MCP server, or a tool that returned nothing useful. A session can be resumed as a live chat from its detail view, and its link shared with anyone who has access to the agent.

​
One pane of glass across the stack

Because the harness runs in the same gateway plane as model and MCP traffic, agent sessions are not a separate silo — they inherit the full AI Gateway observability surface:

Analytics

Model, MCP, and agent traffic in one analytics view, filterable by user, team, and agent.

Request logging

Full request and response logs for every model and tool call, subject to your redaction policy.

Export to OpenTelemetry

Stream traces and metrics to your own observability backend.

Prometheus and Grafana

Scrape gateway metrics and build dashboards and alerts.
For programmatic access to aggregated agent metrics, see Agent Metrics.